Genetic parameters of measured and milk-predicted dry matter intake in Holstein cows

DMI is a key component of feed efficiency in dairy cattle, yet direct measurement requires costly, specialized equipment that is rarely available on commercial farms. The aim of this study was to develop prediction equations for DMI using routinely collected milk information represented either by mid-infrared spectral (MIRS) data or by conventional milk, fat, and protein yields (MFP), and to estimate genetic parameters for measured and predicted DMI in primiparous Canadian Holsteins. Weekly average DMI records (n = 21,363) were available from 1,521 cows in 3 herds. Partial least squares regression (PLSR) models were fitted separately to predict DMI for 2 lactation periods (5-60 and 61-305 d in milk) using either milk MIRS components (512 wavenumbers) or MFP. All predictor sets also included days in milk and age of calving. In early lactation, MIRS-based predictions showed lower animal-level testing-set error than MFP-based predictions, with mean absolute error (MAE) = 2.15 ± 0.08 kg/d and a correlation (r) = 0.70 ± 0.02. In mid-to-late lactation, MIRS also showed slightly lower animal-level testing-set error than MFP (MAE = 2.48 ± 0.08 kg/d; r = 0.61 ± 0.02). These descriptive results suggest that milk MIRS may contain predictive information not fully captured by conventional milk production traits, particularly in early lactation. Held-out herd validation resulted in low to moderate prediction performance for MIRS-based models, whereas temporal validation using the most recent birth year cohort as the testing set yielded stronger prediction performance in both lactation periods (r ≥0.65; MAE ≤2.61). Heritability estimates for MFP-based predicted DMI were moderate (0.27 ± 0.05 in early lactation and 0.36 ± 0.05 in mid-to-late lactation) and comparable to, or higher than, those for measured DMI (0.26 ± 0.05 and 0.27 ± 0.04, respectively), whereas heritabilities for MIRS-based predicted DMI were lower (0.17 ± 0.04 and 0.21 ± 0.04, respectively). MFP-predicted DMI had a moderate genetic correlation with measured DMI (0.52), but also very strong positive genetic correlation with milk production traits (up to 0.99), suggesting that these predictions largely reflect production-related patterns. In contrast, MIRS-predicted DMI showed moderate genetic correlations with measured DMI (0.46-0.55), while showing weak or near-zero correlations with milk production traits. This suggests that MIRS-based predictions may capture additional biological signals of feed intake that are only weakly related to production. Future research should assess the suitability of these proxies for selection by comparing breeding values derived from predicted DMI with official sire evaluations for measured DMI.